AIによるデータ駆動型研究が拓く創薬と医療Data-drivendrugdiscoveryandhealthcarebyAI

山西芳裕1,2YoshihiroYamanishi1,2

1九州工業大学 大学院情報工学研究院FacultyofComputerScienceandSystemsEngineering,KyushuInsCtuteofTechnology2SchoolofPhysicalandMathemaCcalSciences,NanyangTechnologicalUniversity(NTU),Singapore

Outline

•  ドラッグリポジショニング Drug repositioning •  AI創薬・医療 AI-based drug discovery and medicine •  漢方薬リポジショニング Natural medicines repositioning •  再生医療応用 Regenerative medicine

創薬の問題Problemofdrugdiscovery

•  Timeconsuming(10-15years)andhighcost:about1billion$perdrug

•  Highrisk:resultinfailure –  Insufficient efficacy – Unexpected serious toxicity

•  開発コストは高い 数千億円、10年以上 •  ほとんどが失敗に終わる

-  有効性が不十分 -  想定外の深刻な毒性

ドラッグリポジショニング DrugreposiConing

•  IdenCficaConofnewtherapeuCceffects(i.e.,newapplicablediseases)ofexisCngdrugs.

•  Fastdevelopmentandlowrisk(safetyisconfirmed).Example: Sildenafil (Viagra)

Angina → Erectile dysfunction → Pulmonary hypertension

•  既存薬の新しい効能を発見し、別の疾患の治療薬として開発

•  高速・低コスト・低リスク（安全性が確認されている）

シルデナフィル（バイアグラ）：狭心症治療薬　→　男性機能障害薬　→　肺高血圧症薬

例

新薬の多くが既存薬の新効果発見でもたらされてきたManydrugshavebeenprovidedbyfindingneweffects

•  ミノキシジル　Minoxidil–  高血圧薬　Hypertensionmedicine→発毛薬　Hairgrowingagent

•  ビマトプロスト　Bimatoprost–  緑内障薬　Glaucomamedicine→まつげを伸ばす薬　EyelashstretchcosmeCc

•  ブプロピオン　Bupropion–  抗うつ剤　AnCdepressant→禁煙補助剤　AdjuvantforsmokingcessaCon

•  レバミピド　Rebamipide–  胃薬　Stomachmedicine→ドライアイの目薬　Eyedropsofdryeye

問題：これまでは偶然の発見に大きく依存していた The previous approach has been dependent on serendipity.

Machine learning methods to predict new associations between drugs and diseases

Drug 1 Disease 1

Disease 2

Disease 3

knowneffects

new effects to be predicted

Drug 2Drug 3

Drug 4

ドラッグリポジショニングのAI創薬 AI-based drug repositioning

薬と疾患の関連を自動的に予測する機械学習の手法を開発

normal patient

Biological system

gene 1

gene 2

gene 3

多様な疾患の分子レベルでの理解が進んできたMolecularunderstandingofavarietyofdiseases

n  disease-causinggenesn  environmentalfactorsn  abnormalgeneexpression

n  病因遺伝子n  環境因子n  発現異常遺伝子・蛋白

疾患の病態を表す分子的な特徴は、異なる疾患間でも共通する場合がある

CharacterisCcmolecularfeaturesareo[ensharedamongdifferentdiseases

common features

disease A disease B

例えば、男性機能障害と肺高血圧症で、PDE5の異常発現は共通していた. For example, the abnormal expression of PDE5 is observed in erectile dysfunction and pulmonary hypertension.

機械学習による治療薬の予測 Drug prediction by machine learning

Predictive models output

Drug candidates for each disease

Model fA(x)： disease A

applicable

non-applicable

Drug class label

Model fB(x)：disease B

Model fC(x)：disease C

Input Chemical structures and target protein profiles of drugs

φ(x)=

1010!1

⎛

⎝

⎜⎜⎜⎜⎜⎜⎜

⎞

⎠

⎟⎟⎟⎟⎟⎟⎟

(Swada et al, J Chem Inf Model, 55(12), 2717–2730, 2015; Sawada et al, Sci Rep, 8:156, 2018)

安価で安全な抗がん剤を開発DevelopmentofanC-cancerdrugswith

lowpricesandlowtoxicside-effects

•  問題 Problem– がんは死亡原因の第1位– 化学療法は重篤な副作用を伴い、薬価は高騰– Cancerisaleadingcauseofdeathworldwide– Cancertreatmentispainfulandexpensive

•  狙い Aim–  ヒトでの安全性が確認されている薬物から抗がん作用薬を新

しく同定する。–  IdenCficaConofnewanCcancereffectsfromexisCngdrugsthathavebeenconfirmedtobesafeforhumans.

Collaboration with Prof. Tani (University of Tokyo) (Iwata et al, J Med Chem, 61, 9583−9595, 2018)

Traditional approachSearch for drugs that

regulate a single biomolecule

biomolecule interaction

Problem:

Molecular interactions are not taken into account

Proposed approachSearch for drugs that regulate a pathway

biomolecule interaction

Solution: Molecular interactions are considered by using pathway information

Targeting a pathway Drug candidate Targeting a single biomoleculeDrug candidate

Pathway-baseddrugdiscovery(Iwata et al, J Med Chem, 61, 9583−9595, 2018)

biomolecule Interaction

Integration of omics data analysis and molecular network analysis

Molecular interaction network(signaling pathways, metabolic pathways, gene regulations, PPIs)

Drug-induced gene expression data

16268drugs

22276 genes

77 cells

prediction

Drug candidates with expected effects

Pathway-baseddrugdiscoveryforcancersExploring drugs that regulate the following pathways: •  Inactivate cell cycle pathways •  Activate p53 signaling pathways •  Activate apoptosis pathways

Collaboration with Prof. Tani (University of Tokyo)

(Iwata et al, J Med Chem, 61, 9583−9595, 2018)

Outline

•  ドラッグリポジショニング Drug repositioning •  AI創薬・医療 AI-based drug discovery and medicine •  漢方薬リポジショニング Natural medicines repositioning •  再生医療応用 Regenerative medicine

＋

Ordinary drug Kampo drug

漢方薬医療　Naturalmedicine

(e.g.,Herbalmedicines,KampodrugsinJapan)

•  Itispopularanduseful,butthemechanismisunclear.

•  身近で有用だが、メカニズムは大半が不明。

OrdinarydrugThemode-of-acCon:onecompound-onetargetinteracCons

Themode-of-acCon:mulCplecompound-mulCpletargetinteracCons

Efficacy

タンパク質compound

EfficacyviacomplexinteracCons

Compound1

Compound2

Compound3

漢方は多くの化合物で構成され、メカニズムは複雑EachKampodrugcontainsmanycompoundsand

themode-of-acConiscomplicated

protein

protein

ProteinA

PorteinB

ProteinC

Kampodrug

OrdinarydrugThemode-of-acCon:onecompound-onetargetinteracCons

KampodrugThemode-of-acCon:mulCplecompound-mulCpletargetinteracCons

Efficacy

タンパク質compound

EfficacyviacomplexinteracCons

Compound1

Compound2

Compound3

protein

protein

ProteinA

PorteinB

ProteinC

Most targets are unknown

漢方は多くの化合物で構成され、メカニズムは複雑EachKampodrugcontainsmanycompoundsand

themode-of-acConiscomplicated

Compounds

Proteins

：interaction pairs：non-interaction pairs

Kampo A

Kampo C

Kampo B

Chemical structure-based prediction by learning millions of known compound-protein interactions

Multiple compounds – Multiple proteins

DiseaseX

DiseaseY

DiseaseZ

Indication prediction

Interaction prediction

Grouping of target proteins

漢方薬ごとに、標的タンパク質群を予測＆グループ化し、

効能予測TargetproteinsandindicaConsforeachKampowerepredicted

(Sawada et al, Sci Rep, 8:11216, 2018; Douke et al, submitted)

漢方薬の標的タンパク質や適応可能疾患を予測PredicConofpotenCaltargetproteinsandnewapplicablediseasesofKampodrugs

Example: “Boiogito”

肥満症

Protein name Protein function

Applicable disease candidate

糖尿病

既知の効能のメカニズムを示唆SuggesConofthemechanismofknownindicaCon

新しい効能を予測NewlypredictedindicaCon

防已黄耆湯（ぼういおうぎとう）の例

Outline

•  ドラッグリポジショニング Drug repositioning •  AI創薬・医療 AI-based drug discovery and medicine •  漢方薬リポジショニング Natural medicines repositioning •  再生医療応用 Regenerative medicine

iPS cell

hepatocyte (liver cell)fibroblast (skin cell)

Direct reprogramming (direct cell conversion)

Previous approach: gene induction

Proposed approach: compound treatment

再生医療への応用　Regenerative medicine 低分子化合物（薬など）によるダイレクトリプログラミング（細胞直接変換）のための情報技術を開発 Computational direct reprogramming (direct cell conversion) by small compounds (e.g., drugs)

(Sekiya and Suzuki, Nature, 475:390-393 2011)

肝細胞皮膚線維芽細胞

これまでの方法： 遺伝子導入による誘導

本研究の方法： 低分子化合物による誘導

Aims: Avoid cancer risk problem caused by viruses

狙い： ウィルスに起因する発がん リスクの問題を回避する

まとめ　Summary•  機械学習によるビッグデータ解析で、薬や化合物

セットの新しい効能の予測が可能。– 治療効果、健康効果– 細胞分化誘導能

•  MachinelearningenablestopredictnewtherapeuCceffectsofdrugcandidatecompounds.– Applicable diseases and health effects – Cell differentiation abilities